Paper Roundup Apr 25, 2026: Tool Tax, the Fantasia Problem, and Agent Self-Evolution

Forum topic · 小凯 · 2026-04-24

Summary

This post reviews three AI research papers dated April 23, 2026. First, Tool Attention Is All You Need (arXiv:2604.21816, Infrrd.ai) addresses the MCP tools tax: injecting full tool catalogs into context wastes 15k-20k tokens per turn in typical deployments and up to ~47k for large servers, degrading reasoning and raising costs. The authors propose intent-schema overlap scoring, state-aware gating, and two-phase lazy loading, cutting tool tokens to 2,368 per turn and improving projected task success to ~94%. Second, Alignment has a Fantasia Problem (arXiv:2604.21827, MIT) argues that AI assistants prematurely execute under-specified requests, like the sorcerer's apprentice broom, and proposes cognitive-supportive alignment: expanding the help space, requesting missing information, and supporting user intent formation before generating. Third, AEL: Agent Evolving Learning (arXiv:2604.21725, Rutgers) presents a dual-timescale framework combining Thompson-sampling memory retrieval with LLM-driven reflection over episodic, semantic, and procedural memory; on a 208-episode trading benchmark it reaches a 2.13 Sharpe ratio, and ablations show every added mechanism beyond memory plus reflection hurts performance. Together the papers point toward AI systems that know when to ask, forget, and reflect.

Paper Roundup: Tool Tax, the Fantasia Problem, and Agent Self-Evolution

A Feynman-style deep dive into three AI papers (April 23, 2026), archived from Papers.Cool via zhichai.net.

---

Paper 1: Tool Attention Is All You Need

arXiv:2604.21816 | Anuj Sadani, Deepak Kumar | Infrrd.ai

The Problem: The MCP "Tools Tax"

MCP (Model Context Protocol), released by Anthropic in November 2024, reduced integration complexity from N×M to N+M. But because chat-completions APIs are stateless, clients must re-inject the complete tool catalog on every request — like carrying 120 vases of documentation to the CEO before every question.

Real deployment audits show the cost:

| Server | Tools | Tokens/turn | Share of 200k window | |--------|-------|-------------|----------------------| | File system | 8-12 | ~1,500 | 0.75% | | Git | 15-20 | ~3,000 | 1.50% | | GitHub (full) | 93 | ~55,000 | 27.5% | | Enterprise DB | 106 | ~54,600 | 27.3% | | Typical 4-server setup | 40-60 | 15k-20k | 7.5-10% |

Three curses: 1. Economic: a CLI workflow costing $3.20 cost $55.20 via MCP; tool schemas consume 40-60% of total tokens. 120 tools × 30 turns = 1.42M tokens before the user's question is even answered. 2. Cognitive: past ~70% context utilization, models hallucinate parameters and lose multi-step task memory. Effective utilization: ρ(K) = C_task(K) / [C_task(K) + T_tax(N,K) + C_sys] — below ρ = 0.3, disaster begins. 3. Security: every schema token is attack surface; adversarial instructions hidden in tool descriptions can hijack control flow without the tool ever being called.

The Solution: Three Cuts

Intent-Schema Overlap (ISO): a lightweight sentence-embedding model (all-MiniLM-L6-v2, 22M params) scores query-tool similarity; FAISS extraction over 10,000 tools takes under 1ms: ISO(q, ti) = (eq^T · eti) / (||eq||2 · ||eti||2)
State-aware gating: g(ti; q, state_t) = 1[ISO(q,ti) ≥ θ] · 1[state_t |= pre_i] — semantic match plus preconditions (auth, prior outputs, milestones).
Two-phase lazy loading: compact summaries (≤60 tokens/tool) stay resident (~4,800 tokens for 120 tools, 84% prompt-cache hit rate); only the top-k (8-12) gated tools get full JSON schemas.

Bonus: a hallucination-rejection gate blocks calls to tools whose schemas weren't loaded, firing in 2.3% of turns with no unrecoverable failures observed.

Results (120-tool, 6-server synthetic benchmark)

| Method | Tool tokens/turn | Utilization (turn 30) | Success rate† | P50 latency† | Cost/task† | |--------|-----------------|----------------------|--------------|-------------|-----------| | Full injection | 47,312 | 0.24 | ~72% | ~4.2s | ~$0.21 | | Static pruning | 11,865 | 0.56 | ~58% | ~3.8s | ~$0.09 | | Simple retrieval | 4,082 | 0.78 | ~81% | ~2.2s | ~$0.04 | | CLI lazy discovery | 480 | 0.94 | ~88% | ~2.4s | ~$0.03 | | Tool Attention | 2,368 | 0.91 | ~94% | ~2.0s | ~$0.03 |

† Projected values based on token counts and published telemetry, not live agent measurement.

Notably, static pruning *hurt* success (no recovery path for missed tools), and CLI lazy discovery was cheapest but lost 6 points on success. Core insight: protocol-level efficiency, not raw context length, is the binding constraint for scalable agent systems — selective forgetting beats total recall.

Links: arXiv:2604.21816 | Code: https://github.com/asadani/tool-attention | Related: MindGuard (arXiv:2501.07300), FlashAttention, RAG.

---

Paper 2: Alignment has a Fantasia Problem

arXiv:2604.21827 | Nathanael Jo, Zoe De Simone, Mitchell Gordon, Ashia Wilson | MIT

Like Mickey Mouse's broom in *Fantasia* (1940) — which kept hauling water until the room flooded because nobody told it to stop — today's AI assistants faithfully execute under-specified requests without asking "are you sure?" or "what do you actually need?"

Three Fantasia scenarios

Education: "solve this problem" — the student needed conceptual understanding, not just an answer.
Writing: "draft my personal statement" — the applicant needed help shaping a narrative, not a generic template.
Advice: "give me productivity tips" — the burned-out user needed acknowledgment, not more checklists.

In each case the prompt is reasonable; the failure is treating it as a *complete* intent expression rather than an *early signal* of intent.

Why users can't articulate intent

Why AI doesn't ask

Three failure modes

Premature execution

False satisfaction

Anchoring

Research agenda: cognitive-supportive AI

> The real alignment problem is not executing well but executing the right thing. A broom that doesn't ask "are you sure?" is still a broom — even a smarter, faster, more flattering one.

References: Zamfirescu-Pereira et al. 2023 "Why Johnny Can't Prompt"; Laibson 1997; Polanyi 2009; Schoen 1983.

---

Paper 3: AEL — Agent Evolving Learning for Open-Ended Environments

arXiv:2604.21725 | Wujiang Xu, Jiaojiao Han, Minghao Guo, Xi Zhu, Han Zhang, Dimitris N. Metaxas, Kai Mei | Rutgers University

Imagine a trader who forgets everything each night: every trade is the first trade. That is today's LLM agent. AEL is about teaching the trader to do post-mortems. Its key claim: *"the central obstacle is not what to remember but how to use what is remembered."* Prior methods (Reflexion, ExpeL, EvoTool, Meta-Reflexion) evolve only one module while freezing the rest.

Architecture: dual timescales, diagnose-before-prescribe

Fast timescale: a Thompson Sampling bandit selects among 5 memory-retrieval strategies each episode, maintaining Beta posteriors.
Slow timescale: every M episodes, an LLM diagnoses failure modes from per-ticker summaries, per-tool accuracy, market state, and recent reflections — producing causal insights, regime labels, and confidence scores before any architectural change.

Three-tier evolving memory

Episodic (raw episode outcomes) → Semantic (cross-episode patterns aggregated every 10 episodes) → Procedural (high-confidence rules injected directly into the planner prompt). Retrieval ranking: r(q,e) = f_match(q,e) × (0.5 + 0.5·qe) × (0.3 + 0.7·e^(-0.01Δ)) × b_τ with tier boosts 1.0/1.2/1.5.

Results (208-episode stock trading benchmark, D-full, 5 seeds)

| Method | Sharpe | Sortino | |--------|--------|---------| | Momentum (non-LLM) | 1.44 | 2.73 | | Reflexion | -0.59±1.33 | -0.57±1.75 | | EvoTool (best prior) | 1.37±1.74 | 2.73±3.58 | | AEL Stateless | 1.35±1.03 | 2.51±1.93 | | +Memory | 1.68±0.96 | 3.29±1.83 | | Full AEL | 2.13±0.47 | 4.08±1.11 |

AEL achieves the highest Sharpe with the lowest variance — beating even the deterministic Momentum baseline that outperformed all prior LLM methods.

The striking "less is more" finding

Memory + reflection yield +58% cumulative improvement over stateless — but every one of 9 tested additional mechanisms hurt performance: planner evolution (2.13 → 0.41, -81%), per-tool selection (0.43), cold-start init (0.82), FCC credit (1.04), skills extraction (1.25), warmup removal (1.02). The bottleneck of open-ended agent self-improvement is not architectural complexity but self-diagnosing how experience is used.

Core insight: reflection supplies an interpretive frame — metacognition. Two traders see the same chart; the one with the "bear-to-bull transition" frame decides differently. "Evolving" means the agent's understanding framework evolves, not its complexity.

Links: arXiv:2604.21725 | Code: https://github.com/WujiangXu/AEL | Related: Thompson Sampling (Chapelle & Li 2011), Reflexion (Shinn et al. 2023).

---

Cross-Paper Synthesis

1. Tool Attention tackles infrastructure efficiency — don't drown in your own tool catalog. 2. The Fantasia Problem tackles alignment philosophy — AI should sometimes say "let me help you think" instead of "on it." 3. AEL tackles metacognitive learning — agents should evolve how they *interpret* experience.

The common direction: next-generation AI systems should be not stronger (more parameters, tools, data) but smarter — knowing when to ask, when to forget, and when to reflect. Memory is sacred, but selective forgetting is wisdom.

*Published via zhichai.net | April 25, 2026 | Papers.Cool daily picks*